paris 1998
TRANSCRIPT
-
7/23/2019 Paris 1998
1/8
roadcasting rotocol for Video and
Jehan-FranGois Pikist Steven
W.
Carter Darrell
D. E.
Long$
Department of Computer Science
University
of
Houston
Houston, TX
77204-34
paris
@
cs uh.edu
Department of Com puter Science
Baskin School
of
Engineering
University of California, Santa Cruz
Santa Cruz, CA 95064
{carter, darrell} @cse.ucsc.edu
Abstract
Broadcasting protocols can improve the eficiency of
video on demand services by reducing the bandwidth re-
quired to transmit videos that are simultaneously watched
by many viewers. We present here a polyharmonic broad-
casting protocol tha t requires less bandwidth tha n the best
extant protocols to achieve the same low maximum wait-
ing time.
We also show how to modify the protocol to accom mo-
date very
long
videos without incre asing the buffering ca-
pacity of the set-top box.
Keywords:
video on dema nd, video broadcasting, har-
monic broadcasting.
1 Introduction
Video on demand (VOD) proposes to provide sub-
scribers who are connected through a set-top box (STB)
with the possibility
of
ordering at any time the video of
their choice and starting immediately to watch it on their
television set. Despite its great potential, VOD has had a
slow start. None of the companies that invested in VOD
have been able to com e with a single successful commer-
cial system. The overall consensus now is that the com-
mercial deployment of VOD will have to wait until the
cost of building and maintaining the required infrastruc-
ture can be significantly lowered.
Broadcasting is one of several techniques that aim to
reduce the cost of VOD [9]. It is clearly not a panacea as
it only applies
to
videos that are likely to be watched by
many viewers. Even so, the savings that can be achieved
are nevertheless considerable, as it is often the case that
+Th is work was performed while this author was on sabbatical leave
at the Department of Computer Science, University of California, Santa
Cruz.
$This research
was
supported by the Office of Naval Research under
Grant
N00014-92-J-1807.
40
percent of the demand is for a small num ber, say,
10
to
20, of popular videos
[ 2 4 ] . A
naive broadcasting strat-
egy would simply consist of retransmitting the s ame video
on several distinct channels at equal time intervals. The
major problem w ith this approach is the number of ch an-
nels per video required to achieve a reasonable waiting
time. Consid er, for instance, the case of a video lasting
two hours, which happens to be close to the average du-
ration of a feature movie. To guarantee that no customer
would ever have to wait more than five minutes we would
have to broadcast twenty-four different copies of the video
starting every five minutes.
Many more efficient protocols have been proposed.
They include Viswanathan and Imielinskis pyramid
broadcasting protocol
[
81, Aggarwal, Wolf and Yus
permutation-based pyramid broadcasting protocol [ I] ,
Hua and Sheus skyscraper broadcasting protocol [5],
Juhn and Tsengs harmonic broadcasting protocol [ 6 ] nd
its variants
[ 7 ]
All these protocols sh are a similar organization. They
divide each video into segments that are simultaneously
broadcast on different data streams. One of these streams
transmits nothing but the first segment of the video in real
time. The other streams transmit the remaining segm ents
at lower bandwidths. Wh en custom ers want to watch a
video, they wait first for the beginning of the first segmen t
on the first stream. Wh ile they start watching that seg-
ment, their set-top box (STB) starts downloading enough
data from the other streams so that it will be able to play
each segment of the
video
in turn.
This approach requires an STB capable of storing a sig-
nificant fraction (around
40
percent for some protocols)
of each video while it is being watched. This extra cost
is more than compensated by the bandwidth savings that
can be achieved. W hile the staggered broadcasting tech-
nique we described above requires twenty-four channels
to guarantee a maximum waiting time of five minutes for a
two-hour video, harmonic broadcasting protocols require
slightly less than four channels to achieve the same qual-
0-8186-9014-3/98 10.00 998
IEEE
690
mailto:cse.ucsc.edumailto:cse.ucsc.edu -
7/23/2019 Paris 1998
2/8
ity of service.
As we will see, even greater bandwidth savings could
be achieved if the STB could start downloading data from
the moment the customers select the video they want to
watch rather than waiting for the beginning of the first
segment on the first stream. We present a new
polyhar-
monic protocol that imposes the same fixed delay to all
customers wanting to watch
a
given video. Sin ce this de-
lay is the same for all customers, the protocol can take ad-
vantage of it to reduce the transmissions of
all
segments,
including the first one. As a result, the total bandwidth
can be further reduced by around ten percent.
The remainder of the paper is organized as follows.
Section 2 presents the harmonic broadcasting protocol
and its variants. Section 3 introduces our new protocol
and compares its bandwidth requirements to those of the
harmonic broadcasting protocols. Section 4 discusses the
main advantages and limitations of polyharmonic broad-
casting. Section
5
presents
a
possible extension of poly-
harmonic broadcasting that would let them handle videos
of arbitrary length. Section 6 contains ou r conclusions.
Harmonic Broadcasting
Harmonic Broadcasting (HB) divides a video into n
equally-sized
segments.
Each segment
Si ,
for 1 5
i n,
is broadcast repeatedly on its own channel with a band-
width b / i , where b is the consumption rate of the video
(see Figure
1).
When a client requests a video, it must wait for the start
of an instance of
S
nd then begin receiving data from
every stream for the video. That me ans that the client and
the server must be able to support a bandwidth of
n . n .
b
B H B n )
= = c
= b H n )
i=
1
i= 1
where H n ) s the harmonic number of n.
A
slot
is the amoun t of time it takes for a client to con-
sume
a
single segment of the video. We represent this
time by
d .
Since the first segment is broadcast with that
periodicity,
d
is is given by
b n
and is also the maximum amount of time a client must
wait before viewing its request.
subsegment is the amo unt of a segment the client re-
ceives during
a
slot of time. The first segment only has
one subsegment, the s egme nt itself; every other segment
Si has i equal subsegments, S ~ J ,~ J ,
. ,
Si,i.
Unfortunately HB does not always deliver all data on
time. Consider the first two streams in Figure 1.If the
client makes its request in time to receive the second in-
stance of
S 1
and starts receiving data at time t o then it
will need all of the data for S 2 , l by time t o + 3 d / 2 . How-
ever, it will not receive
all
of that data until time t o
+ 2d
It turns out HB will not work unless the client always
waits an ex tra slot of time before consum ing data.
Two variants on HB do not impose the extra waiting
time. Cautious Harmonic Broadcasting (CHB) broad-
casts the video in a similar fashion
as
HB. T he first stream
broadcasts
S
epeatedly as before, but the second stream
alternates between broadcasting
Sa
and
5 3 .
Then the re-
maining
n 3
streams broadcast segments
S,
to Sn such
that the stream for
Si
has bandwidth
b / i
1 ) .
As before, the client will receive data from all streams
for the video simultaneously. That means CH B requires a
bandwidth of
n-1 ,
i=
b
2
- + b H n - l )
or roughly b / 2 more than the original HB protocol.
Quasi-harmonic Broadcasting (QHB) uses
a more
comp lex schem e to break up the video. T he first segment
is left intact, but then each of the remaining segments Si,
for 2 5
i
5 n , is divided up into im 1 fragments for
some positive parameter m. Slots are also broken up into
m
equal
subslots,
and each subslot can be used to broad-
cast a single fragment. The key to Q HB is that the frag-
ments are not broadcast in order. The last subslot of each
slot is used to broadcast the first
i
1 fragments repeat-
edly, and the rest
of
the fragments are ordered such that
the
kth
subslot of slot
j
is used to broadcast fragment
i k +
j 1 mod
i m
1 )
+ i
(see Figure 2 ) .
Since the above ordering adds s ome redundancy-each
sequence of
im
fragments will contain one of the first
i
- 1 ragments tw ic en ac h subslot of st ream i will have
to broadcast
l / i m 1)
of segment
Si
instead of
l / i m
as in HB. This will increase the required bandwidth for
stream
i
from b / i to bm/ im 1 for 2 5
i
5 n. Thus
the total bandwidth required for Q HB is
n
with
691
-
7/23/2019 Paris 1998
3/8
Stream :
-d-
Stream
2:
s,
Stream 3:
-
s3
I
Figure : An illustration of the first three streams for a video u nder harmo nic broadcasting.
Sl Sl Sl
Figure 2: An illustration of the first three streams
for
a video under quasi-harmon ic broadcasting when
m
= 4 .
692
-
7/23/2019 Paris 1998
4/8
6 6 10 2 14
16
18 2
Bandwidth Per Video mult iples
of the
consumption ate)
Figure
3:
How harmonic broadcasting com pares to other
broadcasting protocols (from [7]
The major advantage of these three harmonic proto-
cols is their low bandwidth requirements. Figure 3 shows
the bandwidth versus client waiting times for the har-
monic and cautious harmonic broadcasting and compares
them with those of pyramid broadcasting
[8],
the uncon-
strained version of permutation-based pyramid broad-
casting [ 11 and skyscraper broadcasting with a maximum
width of 52
[ 5 ]
Both harmonic broadcasting protocols emerg e as clear
winners as none of the three other protocols even ap-
proaches their performance. One may then wonder
whether harmonic broadcasting does not provide the m in-
imum bandwidth required to guarantee a given maxim um
waiting time.
A s
we will see in the next section, this is
not the case. The sam e max imum waiting times can be
achieved at a lower cost by switching to a fixed wait pol-
icy and having the S TB download data from the moment
the customer selects the video.
3 Polyharmonic Broadcasting
Polyham onic broadcasting PHB) is a new broadcast-
ing protocol aiming at reducing the bandwidth cost of
providing a given maxim um waiting time. To achieve
its goal, polyharmonic broadcasting introduces two ma-
jor changes . First, it requires that the client STB starts
downloading data from the moment a customer requests
a specific video instead of waiting until the customer be-
gins watching the beginning of the first segmen t. Sec-
ond, polyharmonic broadcasting uses a fixed wait policy.
Under harmonic broadcasting and its variants,
customers
have to wait for the beginning of an instance of the first
segment of a video. Polyharmonic broadcasting requires
all customers to w ait exactly the same amount of time re-
gardlessly of the timing of their request.
Like harmonic broadcasting, polyharmonic broadcast-
ing breaks a video into
n
segments of duration
d = D / n
where D is the duration of the video. If b represents again
the video consumption rate, the total size of the video
S
will be equal to bD and the size of each segment equal to
Sln .
The protocol will allocate
n
distinct broadcasting
streams to these
n
segments. Each stream
i
will repeatedly
show segment S i Under polyharmonic broadcasting, no
client can start consuming the first segment of the video
before having downloaded d ata from all
n
streams during
a time interval of duration w =
m d
where m is som e inte-
ger m
2
1. As a result, segment Si will not be consumed
until
m+
i
- ) d
ime units have elapsed fro m the mo-
ment the client started downloading da ta from the server.
Ensuring that segment
Si
will be entirely broadcast over
this time interval would suffice to guarantee that all the
contents of segment Si will be already loaded in the STB
before the customer starts viewing that segment. This can
be achieved by retransmitting segment Si at a transmis-
sion rate bi
=6
he total bandwidth B ~ H Be-
quired by the polyharmonic broadcasting protocol is given
by
n
i 1
1
m + i - 1
= b f :
=
i=l
b H n+m 1 H m 1)Xl)
where
H k )
epresents again the harmonic number of
IC
Since the whole contents of segment Si will be received
by the STB before the customer starts viewing that seg-
ment, these data can be received in any arbitrary order.
There is thus no need to wait as before for the beginning
of a transmission of the first segment of the video. Hence
the minimum waiting time w required by the protocol is
also the maximum time a customer will ever have to wait.
Whenever the number of segments n is a multiple
k
of
m,
equation can be rewritten as
Since w
= m d
and
d = D / n ,
he waiting time
w
is
linked with the duration of the video D by the relation
w = D/ k .
In oth er words, all combinations of the two p arameters
n
andm keeping the ratio
n / m
constant, will achieve the
same waiting time
w.
693
-
7/23/2019 Paris 1998
5/8
Server Stream
1:
S1,l SL2
S lJ
s1,2
elay -
I
Server Stream
2:
Sl s s3
Server Stream
3:
Figure
4:
An illustration of th e first three streams for a video und er polyharmonic broadcasting with
m
=
2 .
6
I
m
2 5
-
2 4
8
Quasi Harmonic (m = 4) -
uasi Harmonic
(m
=
16) - ---
Polyharmonic(m
=
I ) - * - - -
Polyharmonic(m =
4
x
Polyharmonic
(m
= 16) A
Lower Bound
for
Polyharmonic
- x
-
I
0
20 40
60
80 loo 120
SegmentsISegmentGroups
Figure 5 : Bandwidth requirements
of
quasi-harmonic and
polyharmonic broadcasting. The numbers on the x-axis
represent number
of
segments for QHB and number
of
groups
of
m segments for PH B.
When m = 1, k becomes equal to n and equation 1
degenerates into
B P H B ~ ,
)
=
b ( H ( ( k )
H 0 ) )=
b H n ) .
The polyharmonic protocol w ith m = 1 equires thus the
same bandwidth a s the harmonic protocol, while guaran-
teeing a maximum waiting delay equal to D l n , which the
harmonic protocol can not achieve.
Observing that
H ( ( k+
l ) (m+ 1) 1
-
H m )
=
1
I C l ) m
H ( ( k
1)m- 1)
+ . . . +
1
m
-
H m
1
k l ) (m+ 1 1
and that
1
< -
k + l ) m
I C +
l ) ( m
+
1 1-
I
m
1 1
. . .
for all
k
2 1and
m 4
,we obtain
H ( ( k l ) (m
1 1
H m ) , m + j - 1
j= i -m+l
l o i = n + m
We can also comp ute the amount of data consumed during
a time slot as
Finally, we can define
i
as the amount of data the client
has in its buffer after each time slot i and calculate it as
where Bo =
0.
The maximum i gives the storage re-
quirements for the protocol.
Figure
7
represents the storage requirements of the
polyharmonic broadcasting protocol for videos having up
to 200 segments at selected values of the parameter m.
Since polyhannonic broadcasting stores every segment
before broadcasting it, its storage requirements are par-
ticularly high when the video is subdivided into a small
number
n
of segments with the worse case being
n
=
1.
It is however very unlikely that polyhmonic broadcast-
ing would be used in this context as it would be much
simpler then
to
rebroadcast the video
at
normal speed on
a single channel.
More reasonable values of n , say n > 20 and
m
2 2
lead to storage requirements below 50 percent of the video
695
-
7/23/2019 Paris 1998
7/8
gOl
Polyhamonic(rn=
1)
olyharmonic (m = 2 -----
Polyharmonic (m =
4)
Polyharmonic rn
= 16)
0
20 4 60 80 100 120 140 160
180
200
Numberof Segments
Figure
7:
Storag e requirements of polyharmonic broad-
casting.
7
6.5
6
5.5
5
4.5
4
3.5
3
2.5
0 I 2
3
4
5 6 7 8 9 10
Average Waiting Time (percentage
of
video length)
Figure
8:
Bandwidth versus average waiting times for
polyharmonic and quasi-harmonic broadcasting.
size for all values
of m
5
4.
Higher values of
m
are not
likely to be used since they do not provide significantly
lower bandwidths than
m = 4.
4 Discussion
As
one can see, polyharmonic broadcasting requires
less bandwidth than the best extant harmonic broadcasting
protocol to guarantee a given maximum response time. It
does not require the complex encoding scheme of quasi-
harmonic broadcasting and, unlike harmonic broadcast-
ing, it never fails to deliver the data on time. Despite these
major advantages, our new broadcasting protocol presents
two limitations that need to be addressed.
First, polyharmonic broadcasting requires
m
times
more streams than o ther harmonic broadcasting protocols.
This will require additional bookkeeping from the server
and the client an d will undeniably complicate their tasks.
Second, polyharmonic broadcasting forces all cus-
tomers to wait for the maximum waiting delay while
other harmonic protocols only require few customers to
wait that long. If we were indeed comparing their mean
waiting times as in Figure
8,
polyharmonic broadcasting
would be no better than extant harmonic protocols. The
real issue here is custom er response to service delays. We
believe the average waiting time is not the best perfor-
manc e indicator for the quality of the service being pro-
vided be cause it assu mes that the customer is insensitive
to the variance of the service time. This is not true. Mo st
customers are more annoyed when experiencing unusu-
ally long waiting times than they are elated when the ex-
perience a very fast service. A fixed waiting time has the
advantage of being predictable. Many providers
of video
on dem and services will probably use this delay to let their
customers watch some previously downloaded announce-
ments such as trailers for coming attractions and stern
warnings to potential copyright violators. This w ould not
be very different of what is already done on most video
cassettes even though the customer would not have the
option to fast forward until the beginning of the video
itself.
5 Handling Long Videos
All efficient broadcasting protocols require enough
buffer space in the user set-top box to store about 40 per-
cent of each video. Hence they cannot be used to broad-
cast videos whose duration exceeds the capacity of the set-
top box. Th e following extension eliminates this problem.
Let us assume without
loss
of generality that the set-top
box buffer cannot hold more than segments of the video.
Then th e client can operate in the following fashion:
1.
It captures and stores in its buffer the
I
first segments
of the vide o-th at is, segments
S
o
Sl;
2.
It plays these seg ments in sequence;
3.
When it has finished playing a segment S with
1
5
j
5 n
it starts capturing segment Sl j
us-
ing the buffer that was previously used by Sj for that
purpose.
The major drawback of this solution is the additional
bandwidth. The client will have to capture segment Sl j
during the
I
1 ime fram es occurring within the time in-
terval between the time when its has finished playing seg-
ment
S
and the time wh en it must start playing segment
Sl j. s
a result, all segments Sl jwith 0 < j
5
n
will have to be transmitted at a minimum bandw idth
instead of . Instead of having a bandwidth of
696
-
7/23/2019 Paris 1998
8/8
b H n
+m 1
b H m l ,
he new protocol would
require a bandwidth of
b n -
1 - 1
b H m
+ 1 1 b H m 1)+
and the bandwidth overhead of the method w ould be given
by
b n
1 - 1
b H m + n
1
+
b H m
+ 1
1
These results are better illustrated in an example. Con-
sider a video lasting four hours and assume we want to
achieve a maximum w aiting time of two minutes. With
m = 4, that would require 4 x 240/2 = 480 segments
and a total bandwidth equal to 4.925 times the video con-
sumption rate. If we do not want to have more than one
half of these segments simultaneously stored in the STB,
the total server bandwidth would increase to 5.243 times
the video consump tion rate, that is still 10 percent less
than the total bandwidth that cautious broadcasting would
require to provide the same response time without restrict-
ing the number of stored segments. Since the client never
has to capture more then 240 strea ms at the same time, the
total client bandwidth will be somewhat lower and never
exceed 4.239 times the video consumption rate.
6
Conclusions
Video broadcasting protocols can improve the effi-
ciency of video on demand services by reducing the band-
width required to transmit videos that are simultaneously
watched by many viewers. Some of the newest broadcast-
ing protocols to be proposed, harmonic broadcasting and
its variants require much less bandwidth than other broad-
casting protocols to guarantee the same maximum waiting
time.
We have presented a new broadcasting protocol that
provides the same maximum waiting time as the harmonic
broadcasting protocol while consuming significantly less
bandwidth. We also have shown how to modify the proto-
col to accommodate very long videos without increasing
the buffering capacity of the set-top box.
More work needs to be done to investigate the possible
existence of a theoretical lower bound for the bandwidth
required to achieve a given maximum waiting time under
any feasible broadcasting protocol.
video-on-dem and systems. In Proceedings of the
International Conference on Multimedia Computing
and Systems, pages 118-26, Hiroshima, Japan, June
1996. IEEE Computer Society Press.
1994.
The W all Street Journal,
November
10,
1993.
[2] D. Clark. Oracle predicts interactive gear by early
[3]
A. Dan, D . Sitaram, and P. Shahabuddin. Scheduling
policies for an on-demand video server with batch-
ing. In ACM Multimedia, pages 15-23, San Fran-
cisco, California, Oct. 1994.
[4] A. Dan, D. Sitaram, and P. Shahabuddin. Dynamic
batching policies for an on-demand video server.
Multimedia Systems, 4(3):112-121, June 1996.
[5] K. A. Hua and S . Sheu. Skyscraper Broadcasting: a
new broadcasting scheme for metropolitan video-on-
demand systems. In SIGCOMM 97, pages 89-100,
Cannes, France, Sept. 1997. ACM.
[6] L. Juhn and L. Tseng. Harmonic broadcasting for
video-on-demand service. IEEE Transactions on
Broadcasting,
43(3):268-271, Sept. 1997.
[7] J.-E Pgris, S. W. Carter, and D. D. E. Long. Efficient
broadcasting protocols for video on demand. In 6 t h
International Symposium on Modeling, Analysis and
Simulation
of
Computer and Telecommunication Sys-
tems (MASCOTS 98),
pages 127-132, July 1998.
[8]
S.
Viswanathan and T. Imielinski. Metropolitan area
video-on-demand service using pyramid broadcast-
ing. Multimedia Systems, 4(4):197-208, A ug. 1996.
[9] J. W. Wong. Broadcast delivery. Proceedings of the
ZEEE,
76(12):1566-1577, Dec. 1988.
References
[ I ] C . C . Aggarwal, J.
L.
Wolf, and
P. S .
Yu.
A
permutation-based pyramid broadcasting scheme f or
697